More powerful AI models have typically carried higher usage limits due to their greater compute demand. That is no longer the case.
In June, Anthropic made its Claude Fable 5 available to users, a model so capable that it is described as "Mythos-class," a term used to refer to the Mythos model, deemed so powerful that it cannot be widely released. Then, just last week, OpenAI released its counterpart, GPT-5.6. The capabilities of these models show a real step change in performance compared to their predecessors—perhaps more than ever before.
Since the models require as much compute as they do, they were not offered to all users for free at launch. Rather, both companies made it available to their paid subscribers with slightly different approaches: Anthropic made Fable available for a certain amount of time, originally until July 7, while OpenAI made its mid-tier model, GPT-5.6 Terra, available to free users, reserving the flagship Sol and the lightweight Luna for paid, higher plans with some usage limits.
Sensing the pressure from OpenAI's release, Anthropic extended its promotional access to the Fable 5 model for all paid subscribers through July 19 and kept weekly rate limits for Claude Code 50% higher. Thibault Sottiaux, engineering lead for Codex, responded to the Anthropic announcement, acknowledging the competition and saying, "I think GPT 5.6 is pretty good."
Adding to the competition, OpenAI celebrated its seven million active users of Codex and ChatGPT by adding a banked reset to everyone's account. This lets you reset your usage caps if you hit the limit from doing a lot of work with the models.
Again, the comments were flooded with users celebrating the extension and comparing OpenAI's release to Anthropic's.
The reasoning for this back-and-forth is an effort to capture the business of power users. There is clearly user demand, as evidenced by the need to reset quotas and by the X threads celebrating each limit extension in the announcement post. The Deep View's Jason Hiner ran through his quota this weekend and said he could've used the reset, because he ended up having to buy extra tokens even though he's on the Pro plan.
Though they could be making more money in the short run by charging users, the idea is to keep users using the AI even more, which is especially valuable for AI tools because in that time users are building their memory, and once that is built, users are more likely to stay within the same AI since it's gotten to know them. Same for any agents built, which may be difficult to transfer.
Competition is also coming from other companies, with underdogs such as Meta and SpaceXAI continuing to release their own models focused on reducing token costs. Ultimately, users are the beneficiaries of this competition. They can try out both models and then decide which one they might want to commit to as a subscription. According to Morningstar research, "the US AI sector is already yielding massive revenue from end users, with perhaps $100 billion in AI services revenue in 2025."
Our Deeper View #
Ultimately, each AI company is betting that if it gives users free access to some of its top models, those users will become hooked and keep coming back for more. This is a strategy businesses have used since the beginning of time, like when Costco took a loss on its $1.50 hot dog, betting that the loss leader would drive people into the store to shop. But there is an important nuance here: compute and tokens are extremely finite resources right now. Even if they sound like abstract concepts, every free query still carries a real cost. Whether this bet actually pays off is something we can only measure over the long run, and it will likely become much clearer once these companies IPO, and we can actually analyze their earnings reports. But in the end, it will likely depend on how rapidly AI's unit economics decline.